Maybe coding agents do not win by writing better code first; maybe they win by becoming easier to talk to. That is the uncomfortable read I take from Cognition buying Interaction, the company behind Poke. Cognition, maker of Devin, acquired Interaction on July 23, 2026, in a deal valued in the low nine figures. Poke is a personal AI agent that lives in your texts, and Cognition’s stated direction is clear enough: bring more conversational personality into Devin’s interaction style.
As someone who builds bots with code, prompts, routing layers, memory stores, and too many test chats open at once, I see this less as a vanity UX purchase and more as a signal about where agent competition is heading. The next fight is not only about who can execute the longest task chain. It is about which agent people can tolerate, trust, correct, interrupt, and return to every day.
Personality is not decoration anymore
Bot builders often treat personality like frosting. First we build the tool. Then we add tone. Then we write a system prompt that says “be helpful and concise,” ship it, and pretend we made a product with a voice.
That approach is starting to look thin. If an AI agent is going to sit between a user and meaningful work, its conversational behavior becomes part of the product’s core machinery. Does it ask before acting? Does it explain too much? Does it know when to be brief? Does it recover gracefully after misunderstanding? Does it feel like an assistant, a coworker, a compiler with emojis, or a stranger in your inbox?
Cognition buying Poke suggests that a coding-agent company sees these questions as valuable enough to buy, not merely prompt-tune. That matters. Devin is associated with software work. Poke is associated with a personal AI agent living in texts. Combining those ideas points toward a future where the agent’s skill and the agent’s conversational presence are developed together.
Texting agents teach a different lesson than coding demos
A coding demo rewards visible output: a pull request, a fixed bug, a generated file, a passing test. A texting agent lives in a different rhythm. It has to be available, casual, responsive, and socially readable. It has to fit into the user’s existing communication habits rather than demanding a special workbench every time.
That difference is important for builders. A user may forgive a one-off tool for sounding stiff if it performs a narrow task. A recurring agent gets judged more harshly. Every awkward message becomes friction. Every needless confirmation becomes noise. Every overconfident reply damages the relationship.
This is why conversational ability is becoming a competitive advantage. Not because people want bots with fake charm, but because interaction style controls how much work a user is willing to delegate. If the agent feels confusing, needy, or oblivious, users keep it on a short leash. If it communicates well, users give it more room.
Architecture needs a personality layer
For ai7bot readers building smart bots, the practical lesson is not “add jokes.” The lesson is to treat personality as an architecture concern. In my own builds, I think about the personality layer as a control surface that sits across several parts of the system:
-
Intent handling: The bot should know whether the user is asking for action, advice, status, or clarification.
-
Permissioning: The bot’s voice should change when it is about to take an action versus when it is only suggesting one.
-
Memory use: A personal agent should not sound forgetful about context the user expects it to remember.
-
Error recovery: The moment after a mistake is where personality becomes trust or damage.
-
Channel fit: A text-based agent should not speak like a dashboard, and a coding agent should not always speak like a chatty friend.
This is also where many agent stacks are still immature. We have orchestration graphs, tool calls, retrieval, evals, and deployment patterns. Those pieces matter. But the user experiences all of that through conversation. If the conversation is bad, the architecture feels bad, even when the internals are doing useful work.
Devin with a better social interface is a sharper product idea
Cognition’s move is interesting because Devin’s domain is technical work, yet the acquired asset is about interaction. That pairing says a lot. The company is not only trying to improve what Devin can do; it is trying to improve how Devin feels to work with.
For coding agents, that distinction is huge. Developers do not simply want code emitted into a void. They want to inspect intent, steer implementation, stop risky moves, ask for alternatives, and understand tradeoffs. A coding agent with a stronger conversational style can make that loop less exhausting.
I do not read the low-nine-figure valuation as proof that personality alone wins. Execution still matters. Agents still need to complete tasks. A charming bot that fails is just a more pleasant failure. But a capable agent with weak interaction design will also hit a ceiling, because users will not know when to trust it or how to guide it.
Builders should test tone like they test tools
The builder takeaway is simple: stop treating voice as a final prompt pass. Test it. Put personality into eval scenarios. Measure whether users understand what the agent is doing. Check whether the agent asks too many questions or too few. Review transcripts where users abandon the flow. Look at moments where the bot should have been direct, quieter, warmer, or more cautious.
Cognition buying Poke is a sign that conversational AI personality has moved from “nice to have” into product strategy. For bot builders, that should change priorities. The agent’s personality is not the mascot. It is the API between human intent and machine action.
If Devin becomes more natural to message, correct, and collaborate with because of Poke’s influence, the lesson will be bigger than one acquisition. The agent that talks best may not always be the smartest agent in isolation. It may simply be the one people keep using.
đź•’ Published: